configure: optimistically enable native dynlink on recent 32bit darwin hosts

[gasche]

See camlp4/camlp4#108: native dynlink seems to work correctly on recent
OSX machines (Darwin15, OSX 10.11), even on 32 bit setups. 32bit
setups were broken by OSX toolchain changes on darwin{10,11} (10.6,
10.7) so all 32bit darwin versions above darwin9 were pessimistically
disabled.

(Apparently the user demand for 32bit darwin host support is that
core_kernel assumes 32bit machines, at least when used in combination
with js_of_ocaml).

[damiendoligez]

There is no need to be optimistic, I have virtual machines running under Darwin1[234]. Let me test and report here.

[damiendoligez]

I've tried on Darwin12 to Darwin15 and they all failed with this error:

ld: illegal text-relocoation (direct reference) to (global,weak) _caml_gc_regs in roots.pic.o from _caml_call_gc in i386.pic.o for architecture i386

[xavierleroy]

Something weird is going on. At what point of the build does this error occur?

[damiendoligez]

In make opt when trying to build libasmrun_shared.so.

[gasche]

Gentle ping: the camlp4 users are discussing applying the patch on their opam switches, so it would be interesting to find out whether it is actually safe.

[xavierleroy]

I'm still surprised that @damiendoligez observed an error on caml_gc_regs but not on other global variables that are also accessed in asmrun/i386.S such as caml_last_return_address. The only difference I can see is that caml_gc_regs is declared in C without initialization, making it a "common" variable, while the other are explicitly initialized. An experiment would be to change line 226 of asmrun/roots.c into:

value * caml_gc_regs = NULL;

[xavierleroy]

Concerning the need to avoid "common" variables in MacOS X, see https://caml.inria.fr/mantis/view.php?id=5693#c7798

[mshinwell]

@damiendoligez Could you try the suggested change?

[damiendoligez]

The error message is slightly different:

ld: illegal text-relocation to '_caml_gc_regs' in roots.pic.o from '_caml_call_gc' in i386.pic.o for architecture i386

Then, if I comment out line 156 of i386.S:

        movl    %esp, G(caml_gc_regs)

I get this error instead:

ld: illegal text-relocation to '_caml_bottom_of_stack' in roots.pic.o from '_caml_call_gc' in i386.pic.o for architecture i386

So it's not just caml_gc_regs but presumably all the globals that are used in instructions different from pushl and popl.

[xavierleroy]

@damiendoligez 's observations are consistent with the hypothesis that MacOS x86-32 does NOT support dynamic linking of non-PIC code. In itself I'm not surprised. The mystery is the claim that it works found in the camlp4 github.

[damiendoligez]

Maybe it depends on the Xcode version, although I do have the latest version (8.2.1).

[gasche]

It seems that my general optimism in life does not always match reality. I would vote to close the PR, and wait for a user with an actually reproducible story to maybe restart the discussion.

[damiendoligez]

There is some information in MPR#6900 that may be relevant.